As AI workloads scale, Mitsubishi Electric is betting that power, cooling and compute can no longer be designed in isolation
The data centre industry has spent years talking about compute – is it now becoming obvious that compute is only half the problem and should the question now be “how can we build facilities that actually deliver enough power, remove enough heat and remain stable as AI loads become larger and more dynamic” – this is where Mitsubishi Electric’s strategy gets interesting; they recently unveiled “Chip-to-Grid” reference designs for next-generation NVIDIA AI infrastructure, including Vera Rubin NVL72. The idea is fairly straightforward: stop treating the grid, power distribution, GPUs and cooling systems as separate engineering problems and design them as one architecture – which feels like the direction the industry is heading in.
Is power and cooling becoming the same problem?
Traditionally, power and cooling would often be planned as largely separate infrastructure layers – AI workloads make this harder as higher-density racks create bigger electrical loads while simultaneously creating much larger thermal loads – add in 800VDC distribution, direct liquid cooling, on-site generation and battery storage, and the old model of simply adding another UPS or chiller starts to look increasingly inadequate.
Their reference architecture combines conventional 415/480VAC and 800VDC distribution with direct liquid cooling and air cooling, it also incorporates on-site generation and battery energy storage with the ability to operate as an autonomous microgrid during a grid failure. The important point here isn’t individual technology – It is how to connect them.
The real opportunity may be integration
Mitsubishi Electric is clearly trying to move beyond being viewed simply as a supplier of electrical or cooling equipment – their European data centre business positions critical power, cooling, DCIM and modular infrastructure as parts of a broader solution and their DCIM platform already brings power, cooling and IT infrastructure into a single operational view. Their intention is to broaden the offering into areas including power systems, facility monitoring and security – this matters to operators because AI infrastructure is fast becoming too interconnected for operators to optimise every subsystem independently as everything is beginning to touch everything else.
A more powerful GPU cluster changes the electrical design – the electrical design affects heat generation – the cooling architecture affects rack density and rack density affects the economics of the building.
Are AI data centres moving towards total system engineering?
This is arguably the bigger story – the next competitive advantage in AI infrastructure may not come from having the best individual components – it may come from how effectively the entire facility works as a system, which for operators means the conversation is shifting from “How much power can we get?” to something more complicated: “How do we deliver, distribute, cool, store, monitor and control that power efficiently at AI scale”?
Will the answer be an integrated architecture rather than a collection of products? For an industry building out gigawatt-scale AI facilities, this is worth watching.
